Intraday return predictability in cryptocurrency perpetual futures

R2 Cryptocurrency perpetual futures account for 93% of cryptocurrency trading volume, with daily turnover exceeding $100 billion, yet their intraday return predictability remains unexplored: a notable gap given extensive evidence of predictable daily returns in spot markets. We investigate whether this inefficiency extends to perpetual futures at 1-, 4-, and 8-hour horizons using a broad set of machine learning methods. Out-of-sample R2 statistics are uniformly negative: neither linear models nor flexible nonlinear methods outperform a historical-mean benchmark. Trading strategies based on model forecasts underperform passive buy-and-hold, and once realistic transaction costs are incorporated, virtually all active strategies deliver negative risk-adjusted returns. These results suggest a resolution to the cryptocurrency efficiency puzzle: while spot markets display predictable daily patterns, perpetual futures are informationally efficient at intraday frequencies, with any predictive structure too weak and transient to survive high-frequency arbitrage and the implementation frictions of derivatives trading.